OK, so at the link : https://uloz.to/!yyVrCY2Y8sn1/devil-s-elbow-opus-7z (https://uloz.to/%21yyVrCY2Y8sn1/devil-s-elbow-opus-7z) (change the language to English by clicking at the flag at the right upper side of the web page or just simply click at "Stáhnout pomalu" - it may take some time as the file is 55 MB+ and the download service is free) there is 7zip archive with 5 music files. The tune is "Nick Warren - Devil's Elbow" and is freely available from the author's Soundcloud account. I use this song for encoder testing because it requires quite a high bitrate for artifacts to be inaudible (especially the "gas leakage" kind of sound at the very beginning (repeated several times during the full song lenght) of the track. The content of the 7z file is the original track in WAV, 2 files encoded with libopus 1.1.5 at 32 and 48 kb/s (only with --bitrate 32 or --bitrate 48 option, other settings on default) and 2 files encoded with libopus 1.2 RC1, again at 32 and 48 kb/s with default setting. For the 32 kb/s encoded files, there is noticeable drop of stereo (or large part of frequencies) at the beginning of the track roughly between 0:03 and 0:08 which is actually audible for both libopus 1.1.5 and 1.2 RC1 (I haven't noticed it before), but in the case of 1.2 RC1 it's much stronger and probably a bit longer. I think this "effect" occurs several times again further in the track but those occurences are always very short (in tenths of a second). At 48 kb/s, this is not an issue (at least not very obvious) for libopus 1.1.5 but still happens for libopus 1.2 RC1, although the stereo/frequency drop is shorter than in the case of 32 kb/s. At 64 kb/s (file not included in the 7z archive) I can't hear it for both 1.1.5 and 1.2 RC1. The other thing is that while the general sound quality of 32 kb/s encoded files is clearly higher for libopus 1.2 RC1 than for 1.1.5, at 48 kb/s the sound seems to be a bit clearer and closer to the original when using 1.1.5 than 1.2. This could be my subjective feeling as the difference is really tiny (unlike the bitrate, which is noticeably higher for the same bitrate setting for libopus 1.2 RC1 for every file). Best regards, Matěj -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.xiph.org/pipermail/opus/attachments/20170618/98a8a7a1/attachment.html>
Hi Matěj, I had a look at your files and it seems like the issue is that the speech/music detector thinks that some segments are speech rather than music and at low bitrates SILK tends to significantly narrow the stereo image. We're planning on adding an option to opusenc to force it to treat the input as speech/music. Just to confirm that's the issue, can you try encoding with the following option: --set-ctl 4024=3002 and see if that fixes the issue? Cheers, Jean-Marc On 18/06/17 03:45 PM, Matěj Steklík wrote:> OK, so at the link : https://uloz.to/!yyVrCY2Y8sn1/devil-s-elbow-opus-7z > <https://uloz.to/%21yyVrCY2Y8sn1/devil-s-elbow-opus-7z> (change the > language to English by clicking at the flag at the right upper side of > the web page or just simply click at "Stáhnout pomalu" - it may take > some time as the file is 55 MB+ and the download service is free) there > is 7zip archive with 5 music files. The tune is "Nick Warren - Devil's > Elbow" and is freely available from the author's Soundcloud account. I > use this song for encoder testing because it requires quite a high > bitrate for artifacts to be inaudible (especially the "gas leakage" kind > of sound at the very beginning (repeated several times during the full > song lenght) of the track. The content of the 7z file is the original > track in WAV, 2 files encoded with libopus 1.1.5 at 32 and 48 kb/s (only > with --bitrate 32 or --bitrate 48 option, other settings on default) and > 2 files encoded with libopus 1.2 RC1, again at 32 and 48 kb/s with > default setting. For the 32 kb/s encoded files, there is noticeable drop > of stereo (or large part of frequencies) at the beginning of the track > roughly between 0:03 and 0:08 which is actually audible for both libopus > 1.1.5 and 1.2 RC1 (I haven't noticed it before), but in the case of 1.2 > RC1 it's much stronger and probably a bit longer. I think this "effect" > occurs several times again further in the track but those occurences are > always very short (in tenths of a second). At 48 kb/s, this is not an > issue (at least not very obvious) for libopus 1.1.5 but still happens > for libopus 1.2 RC1, although the stereo/frequency drop is shorter than > in the case of 32 kb/s. At 64 kb/s (file not included in the 7z archive) > I can't hear it for both 1.1.5 and 1.2 RC1. The other thing is that > while the general sound quality of 32 kb/s encoded files is clearly > higher for libopus 1.2 RC1 than for 1.1.5, at 48 kb/s the sound seems to > be a bit clearer and closer to the original when using 1.1.5 than 1.2. > This could be my subjective feeling as the difference is really tiny > (unlike the bitrate, which is noticeably higher for the same bitrate > setting for libopus 1.2 RC1 for every file). > > Best regards, > Matěj > > > _______________________________________________ > opus mailing list > opus at xiph.org > http://lists.xiph.org/mailman/listinfo/opus >
Hello Jean-Marc, yes, it works for both 32 and 48 kb/s. Also this option marginally increases the size of the file which is understandable. I tried a few other music files and found that this stereo narrowing happens not only in the case of track I submitted (although in smaller scale) so it would be great if the final libopus 1.2 had officialy a parameter to force music "mode" of encoding. I'd like to ask two additional questions: 1) is the -- set-ctl 4024=3002 parameter a present alternative to that planned "force music encoding" option or is it an other workaround?; 2) if I wanted to set opusenc to not use SILK at all, is there a minimal bitrate which I have to set (such as 48 or 64 kb/s) or is it that when during VBR encoding the actual bitrate at one point drops below some value, the SILK or hybrid mode is used even if the target file bitrate is in CELT territory (48 kb/s+)? Thanks for answer. Best regards, Matěj ---------- Původní e-mail ---------- Od: Jean-Marc Valin <jmvalin at jmvalin.ca> Komu: Matěj Steklík <matej.steklik at post.cz>, opus at xiph.org Datum: 19. 6. 2017 5:39:15 Předmět: Re: [opus] Stereo dropping to mono with libopus 1.2 RC "Hi Matěj, I had a look at your files and it seems like the issue is that the speech/music detector thinks that some segments are speech rather than music and at low bitrates SILK tends to significantly narrow the stereo image. We're planning on adding an option to opusenc to force it to treat the input as speech/music. Just to confirm that's the issue, can you try encoding with the following option: --set-ctl 4024=3002 and see if that fixes the issue? Cheers, Jean-Marc On 18/06/17 03:45 PM, Matěj Steklík wrote:> OK, so at the link : https://uloz.to/!yyVrCY2Y8sn1/devil-s-elbow-opus-7z > <https://uloz.to/%21yyVrCY2Y8sn1/devil-s-elbow-opus-7z> (change the > language to English by clicking at the flag at the right upper side of > the web page or just simply click at "Stáhnout pomalu" - it may take > some time as the file is 55 MB+ and the download service is free) there > is 7zip archive with 5 music files. The tune is "Nick Warren - Devil's > Elbow" and is freely available from the author's Soundcloud account. I > use this song for encoder testing because it requires quite a high > bitrate for artifacts to be inaudible (especially the "gas leakage" kind > of sound at the very beginning (repeated several times during the full > song lenght) of the track. The content of the 7z file is the original > track in WAV, 2 files encoded with libopus 1.1.5 at 32 and 48 kb/s (only > with --bitrate 32 or --bitrate 48 option, other settings on default) and > 2 files encoded with libopus 1.2 RC1, again at 32 and 48 kb/s with > default setting. For the 32 kb/s encoded files, there is noticeable drop > of stereo (or large part of frequencies) at the beginning of the track > roughly between 0:03 and 0:08 which is actually audible for both libopus > 1.1.5 and 1.2 RC1 (I haven't noticed it before), but in the case of 1.2 > RC1 it's much stronger and probably a bit longer. I think this "effect" > occurs several times again further in the track but those occurences are > always very short (in tenths of a second). At 48 kb/s, this is not an > issue (at least not very obvious) for libopus 1.1.5 but still happens > for libopus 1.2 RC1, although the stereo/frequency drop is shorter than > in the case of 32 kb/s. At 64 kb/s (file not included in the 7z archive) > I can't hear it for both 1.1.5 and 1.2 RC1. The other thing is that > while the general sound quality of 32 kb/s encoded files is clearly > higher for libopus 1.2 RC1 than for 1.1.5, at 48 kb/s the sound seems to > be a bit clearer and closer to the original when using 1.1.5 than 1.2. > This could be my subjective feeling as the difference is really tiny > (unlike the bitrate, which is noticeably higher for the same bitrate > setting for libopus 1.2 RC1 for every file). > > Best regards, > Matěj > > > _______________________________________________ > opus mailing list > opus at xiph.org > http://lists.xiph.org/mailman/listinfo/opus >" -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.xiph.org/pipermail/opus/attachments/20170619/b50c1bd0/attachment.html>